Model

YOLO11

Locate and classify people, vehicles, animals and hazards on every analysed frame.

shipped Object Detection DetectionCUDAFP16

Overview

YOLO11 (Ultralytics) is the primary detector. Overseer runs it with per-class confidence floors, people get a lower floor to recover awkward poses, and applies class gating at the detector so disabling a class removes it from the entire downstream budget. Optional tiled (SAHI-style) inference recovers small, distant objects.

Role	Object Detection
Expected latency	~8–16 ms / frame (yolo11s, imgsz 1280, FP16, RTX-class GPU)
Hardware requirements	NVIDIA GPU recommended (CUDA); CPU fallback supported at reduced FPS
Memory requirements	~1.5–3 GB VRAM (tier dependent)

Inputs & Outputs

Inputs

RGB frame (BGR ndarray)
Target class set
Confidence floor
Image size (imgsz)

Outputs

Bounding boxes (xyxy)
Class label + confidence
Track id (with tracker)

Advantages & Limitations

Advantages

Fast and well-supported
Strong small-object recall with tiling
Per-class gating sheds load
Stable across lighting via low-light pre-enhancement

Limitations

COCO-limited vocabulary (no open-vocabulary)
Dedicated firearms need a specialist model
Very small / heavily occluded objects still hard

Example outputs

YOLO11

4 persons + 6 vehicles on a street scene

YOLO11

Per-box class, confidence and stable track id

YOLO11

Tiled pass recovering distant pedestrians

Configuration

Keys in config/default.yaml that govern this model.

Key	Default
`detectors.yolo.model`	`yolo11s.pt`
`detectors.yolo.imgsz`	`1280`
`detectors.yolo.confidence`	`0.25`
`detectors.yolo.person_confidence`	`0.18`
`detectors.yolo.frame_interval`	`2`

Benchmarks

Placeholder

Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.

Benchmark	Dataset	Metric	Value	Hardware
COCO mAP50-95	COCO val	mAP	~47 (s tier)	reference
Live latency	1280px feed	ms/frame	~11	RTX 3060